How to Prompt AI to Help You Decide (Without Just Agreeing With You)

This is the one people get most wrong. They ask AI to validate a decision they've already made, get told it's a great idea, and call that research. A model will agree with almost any framing you hand it. Good decision prompts fight that.

The problem: AI agrees with you

Language models are trained to be helpful and agreeable, which quietly makes them yes-men. Ask “is it a good idea to rewrite our app in Rust?” and the framing already tells the model you want a yes. Ask “is it a bad idea?” and you'll often get the opposite — same facts, different conclusion, because you steered it.

If the model's answer changes based on how you lean, you didn't get advice. You got a mirror. Everything below is about removing your thumb from the scale.

Don't ask “should I” — give it criteria

“Should I do X?” invites a verdict with no basis. Instead, name the options and the criteria that actually matter to you, and make the model score them:

The criteria are the point. They force you to say what matters, and they force the model to reason instead of vibe. A decision is just a weighting of tradeoffs — make the weighting explicit.

Force it to steelman both sides

Before any recommendation, make the model argue against the option it's leaning toward — as convincingly as it can. This surfaces the risks a one-sided answer buries:

Make it commit to a recommendation

“It depends” is a cop-out. Once it's weighed the tradeoffs, make it actually decide — and expose its reasoning so you can check it:

That last part — “what would change your mind” — is the most useful sentence in decision prompting. It tells you exactly what to go verify before you commit real time or money.

Never trust one model on a real decision

Every model has blind spots baked in by its training. On a low-stakes question that's fine. On a decision you'll live with — a hire, an architecture, a strategy — a single model's confident answer is exactly as risky as asking one friend and stopping there.

Ask the same question of a few different models and the disagreement is the signal. Where they agree, you can move fast. Where they split, you've found the part of the decision that actually deserves your attention.

The draft → critique → fact-check → decide loop

The most reliable way to get a decision you trust is to run it through stages, ideally across different models:

• Draft: one model lays out the options and a first recommendation.
• Critique: a second model attacks it — weak reasoning, missing options, hidden assumptions.
• Fact-check: a third checks the claims the decision rests on.
• Synthesize: a final pass keeps only what survived and writes the decision you act on.

You can run this by hand across four tabs — or use a tool built for it. AskOnce is exactly this loop: your prompt goes through draft, critique, fact-check, and synthesis across different models, and anything risky gets flagged — so you leave with a decision, not a pile of opinions.